Phison is reimagining the role of the SSD by moving processing power directly onto the storage device. The company recently detailed a new computational-storage architecture that integrates a Neural Processing Unit (NPU) and dedicated local memory alongside NAND flash. By bringing compute to the data, Phison claims the design can slash data transfer requirements by 10 to 100 times for specific workloads, effectively bypassing the traditional bottlenecks of moving massive datasets to a central CPU or GPU.
Processing power at the edge of storage
The heart of this architecture is an NPU capable of 36,800 DMIPS, supported by 64GB of attached memory featuring 130GB/s of bandwidth. This setup is designed to handle modern AI and machine learning frameworks natively, with Phison targeting support for PyTorch, vLLM, OpenVINO, and llama.cpp. While the computational SSD remains in development without a formal product name or release window, it signals a shift toward autonomous storage components that can filter and process information before it ever leaves the drive.
Next-generation speeds and mixed-media efficiency
Alongside its AI-focused research, Phison is pushing the limits of raw throughput with its upcoming PCIe 6.0 controller, dubbed the X3 or PS5303. Leveraging the 64 GT/s lane transfer rate of the Gen6 standard, the X3 aims for sequential speeds up to 28GB/s and 6.8 million IOPS. To support this ecosystem, the company is also developing a suite of infrastructure hardware, including redrivers and retimers.
Phison is further optimizing storage density through a "Mixed Media" design. This approach combines high-speed pSLC flash and high-density QLC flash behind a single 16-channel controller. By integrating fast cache and dense storage within the same drive, the design seeks to balance performance and capacity for data-heavy environments. While these technologies represent the future of the company's roadmap, Phison has yet to provide specific sampling schedules for the new hardware.